High-purity hydrogen fluoride condensation parameter optimization method

Abstract

The invention relates to the technical field of fluorine chemical industry, and discloses a high-purity hydrogen fluoride condensation parameter optimization method. The method comprises the following steps: receiving real-time monitoring data in a condensation process, analyzing a condensation optimization intention by utilizing a physical parameter analysis model and combining a temperature parameter, a pressure parameter and a historical operation record, and generating a parameter optimization vector; the method comprises the following steps: constructing a condensation state index map for operation state data stored in a distributed manner, representing condensation state points through nodes, representing a physical conversion relationship between the state points through edges, and updating a map structure and index efficiency in real time based on an adaptive adjustment mechanism; according to the parameter optimization vector, dynamically positioning an adjustment domain related to the optimization intention in the condensation state index map in a multi-level mapping calculation mode; and in the regulation domain, a candidate parameter set conforming to the optimization intention is screened by using a multi-dimensional fusion matching algorithm, and is sorted and output. According to the method, pertinence and adaptability of condensation parameter optimization are improved, and the high-purity hydrogen fluoride condensation requirement under the complex working condition can be effectively met.

Description

Technical Field

[0001] The present invention relates to the technical field of fluorine chemical industry, and in particular to a method for optimizing condensation parameters of high-purity hydrogen fluoride. Background Art

[0002] High-purity hydrogen fluoride (HF) is a key building block for high-tech industries such as electronics and information technology and photovoltaic energy. Its purity directly impacts the performance and quality of downstream products. In the production of HF, condensation is a core step in achieving gas-liquid separation and removing impurities. Precisely controlling condensation parameters can effectively improve product purity and yield. However, the condensation process is dynamically influenced by multiple factors, including feed gas composition, ambient temperature, and system pressure, posing numerous challenges to parameter optimization. Adjusting parameters for high-purity hydrogen fluoride condensation often relies on operator experience or simple PID control, making it difficult to cope with complex and changing operating conditions. Traditional methods underutilize real-time monitoring data from the condensation process, often adjusting parameters based solely on a single temperature or pressure indicator. This approach lacks in-depth analysis of multi-parameter relationships, leading to a disconnect between optimization intent and actual operating requirements. Furthermore, the condensation system's operational status data is stored in a decentralized manner across different devices or modules, with inconsistent data formats, making it difficult to fully represent the status and making it difficult for operators to quickly grasp the overall system operating status. The description of the condensation state mostly adopts a static model, which cannot reflect the physical conversion relationship between state points in real time. When the raw gas flow, components or environmental conditions fluctuate, it is difficult for the static model to dynamically update the state representation, resulting in the parameter adjustment lagging behind the operating conditions. In addition, in the parameter optimization process, the positioning of the adjustment domain lacks a scientific calculation method, and often relies on experience to define the range, which is prone to over-adjustment or under-adjustment, affecting the condensation efficiency. At the same time, parameter screening mostly adopts a single-dimensional matching method, ignoring the coupling effect between multiple parameters such as temperature, pressure, and flow, making it difficult to screen out the optimal parameter combination that is truly suitable for the current operating conditions. These problems make the high-purity hydrogen fluoride condensation process less stable and the product purity fluctuates greatly, which cannot meet the stringent requirements of high-end industries for high-purity hydrogen fluoride. With the rapid development of high-tech industries, purity requirements for high-purity hydrogen fluoride are constantly increasing. Traditional condensation parameter optimization methods are no longer able to meet the high-precision and high-stability requirements of industrial production. Real-time perception, dynamic modeling, and precise parameter adjustment of the condensation process have become key issues in improving the quality of high-purity hydrogen fluoride production. Summary of the Invention

[0003] The object of the present invention is to provide a method for optimizing the condensation parameters of high-purity hydrogen fluoride to solve the problems raised in the above background technology.

[0004] To achieve the above object, the present invention provides a method for optimizing high-purity hydrogen fluoride condensation parameters, the method comprising: receiving real-time monitoring data of the condensation process, analyzing the condensation optimization intention based on temperature parameters, pressure parameters and historical operation records in the real-time monitoring data using a physical parameter analysis model, and generating a parameter optimization vector based on the condensation optimization intention; A condensation state index graph is constructed for the distributed storage of operational state data in the condensation system. The graph represents condensation state points through nodes, and the edges represent the physical conversion relationships between state points. The graph structure and index efficiency are updated in real time based on an adaptive adjustment mechanism. According to the parameter optimization vector, dynamically locating the regulation domain related to the optimization intention in the condensation state index map in a multi-level mapping calculation manner; In the regulation domain, a multi-dimensional fusion matching algorithm is used to screen candidate parameter sets that meet the optimization intention and sort and output them.

[0005] Preferably, receiving real-time monitoring data of the condensation process, using a physical parameter analysis model, analyzing the condensation optimization intention according to temperature parameters, pressure parameters and historical operation records in the real-time monitoring data, and generating a parameter optimization vector based on the condensation optimization intention, includes: Mapping real-time monitoring data into a fixed-dimensional feature space through physical modeling to generate temperature feature embedding and pressure feature embedding; Obtain historical operation records, analyze historical temperature fluctuations, historical pressure changes, and operation adjustment behaviors, build operation profiles, and convert operation profile data into operation preference embedding; Access the condensation knowledge base, extract the physical constraint features related to the real-time monitoring data, and generate constraint feature embeddings; An algorithm based on weighted fusion is used to integrate temperature feature embedding, operation preference embedding and constraint feature embedding to perform condensation optimization intention analysis. An optimization weight is assigned to each analyzed intention, and a parameter optimization vector is generated based on the optimization weight. The parameter optimization vector covers the current optimization intention and reflects the system operation characteristics.

[0006] Preferably, a condensation state index graph is constructed for the distributedly stored operation state data in the condensation system, wherein the condensation state index graph represents condensation state points through nodes, and edges represent physical conversion relationships between state points, and the graph structure and index efficiency are updated in real time based on an adaptive adjustment mechanism, including: Signal processing technology is used to extract state point features from the operation state data, each condensation state point is used as a node in the condensation state index graph structure, and corresponding state point features are assigned to each node; Create edges in the condensed state index graph structure according to the physical transformation relationship between state points, and set the initial strength of each edge to obtain the condensed state index graph, where the strength is initialized by comparing the transformation similarity of two state points; Based on an adaptive adjustment mechanism, changes in state point parameters and their conversion relationships are tracked to ensure the real-time performance of the condensed state index graph. When adding or modifying state points, an incremental adjustment strategy is used to update only the affected nodes and edges, and the edge strength is dynamically adjusted based on the operation record. Deploy a real-time condensation state index graph in a distributed system.

[0007] Preferably, dynamically locating the regulation domain related to the optimization intention in the condensation state index map in a multi-level mapping calculation manner according to the parameter optimization vector includes: Use the feature mapping algorithm to pre-train the state mapping of the condensed state index graph to generate the state embedding of each node in the graph; Mapping the parameter optimization vector into a state embedding space using a deep alignment model; Calculate the matching degree between the mapped parameter optimization vector and all state embeddings, and select a specified number of nodes with the highest matching degree from the condensed state index graph as the initial adjustment nodes; Based on the neighboring nodes of each initial adjustment node, the first-level neighbor nodes are expanded outward to determine whether the matching degree of the neighbor nodes reaches a specific threshold. If not, the expansion is stopped. If it reaches, the candidate node set is added and the expansion process is repeated to continue exploring the next-level neighbor nodes until the preset level limit is reached or the cumulative node number condition is met; All candidate node sets obtained through multi-level mapping expansion are determined as the regulation domain.

[0008] Preferably, within the regulation domain, a multi-dimensional fusion matching algorithm is used to screen and sort the candidate parameter sets that meet the optimization intention, and output them in order, including: For each condensation state point contained in the regulation domain and possibly related to the optimization intention, extracting multi-dimensional operation features; Use a multi-dimensional fusion algorithm to fuse the operation features of different dimensions to generate the fused operation features of each state point; Calculate the matching degree between the fusion operation characteristics of each state point and the parameter optimization vector, and select the state points with matching degrees higher than the preset threshold as the candidate parameter set; The candidate parameter sets are sorted according to matching degree, operation preference and parameter priority.

[0009] Preferably, the method further comprises: Obtaining a historical operation sequence of a high-purity hydrogen fluoride condensation process, and dividing the process into a stable operation segment and a variable operation segment according to a parameter change trend in the historical operation sequence; According to the difference between the parameter values ​​in the stable operation section and the variable operation section, the parameter types to be optimized are selected; Under each parameter type to be optimized, based on the data similarity of the operating point in the stable operation segment and the changing operation segment, combined with the time characteristics and parameter adjustment records, the optimization capability index of each operating point in the stable operation segment is obtained; According to the parameter adjustment record and parameter reproduction data of each operating point in the changing operation segment, combined with the optimization capability index, the core node probability of each operating point in the stable operation segment is obtained; Each operating point in the stable operating segment is selected as a core node according to the core node probability, and an optimization chain structure is constructed by combining the operation data under the same parameter type in the changing operating segment. Based on the optimization chain structure, optimization schemes for different parameters are determined.

[0010] Preferably, under each parameter type to be optimized, based on the data similarity of the operating point in the stable operation segment and the changing operation segment, combined with the time characteristics and parameter adjustment records, the optimization capability index of each operating point in the stable operation segment is obtained, including: Under any parameter type to be optimized, all points with operating data in the stable operation segment are selected as candidate operating points, and all points with operating data in the changing operation segment are selected as adjustment points. Record any candidate operating point as the selected candidate point, and any adjustment point as the selected adjustment point; According to the parameter adjustment time and adjustment duration of the selected candidate point and the selected adjustment point, the delay influence intensity of the selected candidate point relative to the selected adjustment point is obtained; Calculate the similarity between the data of each adjustment behavior of the selected candidate point in the stable operation segment and the data of each adjustment behavior of the selected adjustment point in the changing operation segment, and obtain the similarity feature factor corresponding to the selected adjustment point under each adjustment behavior; Taking the delay impact intensity as the weight, the similar characteristic factors corresponding to the selected adjustment point under each adjustment behavior are weighted averaged to obtain the parameter optimization force of the selected candidate point relative to the selected adjustment point, and the average of the parameter optimization forces of the selected candidate point relative to all adjustment points is used as the optimization capability index of the selected candidate point.

[0011] Preferably, the step of obtaining the core node probability of each operating point in the stable operating segment based on the parameter adjustment record and parameter reproduction data of each operating point in the changing operating segment in combination with the optimization capability index includes: According to the adjustment time and decision interval of each adjustment behavior of the selected candidate point, combined with the number of parameter recurrences, the decision-making power index of the selected candidate point is obtained; The product of the optimization capability index and the decision-making power index of the selected candidate point is normalized to generate the core node probability of the selected candidate point.

[0012] Preferably, the step of selecting each operating point in the stable operating segment as a core node according to the core node probability and combining the operating data of the changing operating segment under the same parameter type to construct an optimized chain structure includes: The candidate operation points whose core node probability is greater than the preset probability threshold are respectively used as the core nodes of each chain in the optimized chain structure; Obtain the decision-making power index of each adjustment point, and multiply the delay impact strength corresponding to each adjustment point and the core node by the decision-making power index of the adjustment point as the followability index of each adjustment point relative to the core node; For any core node, a chain structure is constructed in descending order of the followability index of each adjustment point relative to the core node. The followability index of nodes at the same level in the chain structure is the same. The chain structure of all core nodes constitutes an optimized chain structure.

[0013] Preferably, the optimization scheme for determining different parameters based on the optimized chain structure includes: In the optimized chain structure, the first shared child node of different core nodes is located at the level where the first target layer is located; after the first target layer in the optimized chain structure, the level with the largest number of child nodes is used as the second target layer; The preset first optimization strategy is used for the operating points between the core node and the first target layer, the preset second optimization strategy is used for the operating points between the first target layer and the second target layer, and the preset third optimization strategy is used for the operating points between the second target layer and the bottom layer.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This high-purity hydrogen fluoride condensation parameter optimization method provides a more systematic and accurate technical path for condensation parameter optimization through the synergistic effect of multiple links. At the data utilization level, the method receives real-time monitoring data of the condensation process, combines it with the physical parameter analysis model, and deeply integrates real-time parameters such as temperature and pressure with historical operation records. It can more comprehensively capture the working condition characteristics and make the analysis of optimization intentions more in line with actual production needs. This method of generating optimization intentions based on actual data avoids the subjectivity of traditional empirical adjustments, makes the generation of parameter optimization vectors more targeted, and can effectively adapt to the condensation needs under different raw material components and flow fluctuations. At the state representation level, a condensation state index graph is constructed based on the distributed operational state data stored in the condensation system. This intuitively presents condensation state points and their transition relationships through nodes and edges, overcoming the limitations of traditional static models for state description. Furthermore, the introduction of an adaptive adjustment mechanism allows the index graph to update its structure and indexing efficiency in real time as operating conditions change, ensuring the timeliness and accuracy of state representation. This allows operators or control systems to monitor the dynamic changes in the condensation system in real time, clearly identifying key nodes for state transitions and providing a clear state reference framework for subsequent parameter adjustments. At the regulatory domain positioning level, a multi-level mapping calculation method is used based on the parameter optimization vector to dynamically locate the regulatory domain, breaking away from the traditional, extensive model of empirically defining the regulatory range. Through layered focus, the multi-level mapping calculation can precisely delineate the regulatory range relevant to the current optimization intent within the complex state space, reducing the search area for invalid parameters and improving the efficiency of parameter optimization. Furthermore, the dynamic positioning method adjusts the regulatory domain boundaries in real time as operating conditions change, ensuring that the regulatory range remains consistent with actual requirements and avoiding regulatory failures due to fluctuating operating conditions. At the parameter screening level, a multi-dimensional fusion matching algorithm is used to screen and rank candidate parameter sets, changing the one-sided nature of traditional single-dimensional parameter screening. This algorithm comprehensively considers the matching degree of multi-dimensional parameters such as temperature, pressure, and condensation rate with the optimization intent. Through multi-factor weight allocation and collaborative calculation, it ensures that the selected candidate parameter set fully meets the optimization requirements. The sorted output method provides operators or automatic control systems with clear parameter selection priorities, facilitating the rapid determination of the optimal adjustment plan and reducing the cost of parameter trial and error. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a timing diagram of the method for optimizing the condensation parameters of high-purity hydrogen fluoride according to the present invention; Figure 2 Flowchart for building and updating the condensation state index map; Figure 3 Flowchart for dynamic positioning of regulatory domains; Figure 4 Flowchart for parameter optimization scheme based on optimization chain structure. DETAILED DESCRIPTION

[0016] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0017] See also Figure 1The present invention provides a method for optimizing the condensation parameters of high-purity hydrogen fluoride. The method comprises: realizing accurate optimization of the condensation parameters through dynamic analysis of real-time monitoring data and historical operation records, combined with a physical parameter analytical model and a condensation state index map. The specific process includes: The system receives real-time monitoring data from the condensation process, including temperature and pressure parameters. Combined with historical operational records, it uses a physical parameter parsing model to analyze the condensation optimization intent and generate a parameter optimization vector. Based on the distributed storage of operational status data, it constructs a condensation state index graph, with nodes representing condensation state points and edges representing the physical transformation relationships between state points. The graph structure and index efficiency are updated in real time based on an adaptive adjustment mechanism. Based on the parameter optimization vector, a multi-level mapping calculation is used to dynamically locate the adjustment domain within the condensation state index graph. Within the adjustment domain, a multi-dimensional fusion matching algorithm is used to screen and sort the candidate parameter sets, outputting the optimal parameter adjustment solution.

[0018] Example 1: See Figure 2 , the temperature and pressure parameters in the real-time monitoring data are converted into a fixed-dimensional feature expression space through physical modeling technology. The dynamic changes of the temperature parameters at continuous timestamps are mapped into a time series feature vector, which contains the characteristics of fluctuation frequency, gradient trend and steady-state offset. The pressure parameters are decomposed into spatial distribution features through the pressure conduction model to form a multi-dimensional tensor structure. During the feature embedding process, the original data in the time window is sampled in segments, and the feature contribution value of each segment of data is calculated based on the thermodynamic transfer function, and finally integrated into the temperature feature embedding vector and the pressure feature embedding vector. This type of embedding vector has dimensional consistency, which facilitates subsequent fusion processing.

[0019] The analysis of historical operation records covers two dimensions: operational behavior patterns and system response characteristics. Historical temperature fluctuation data is divided into low fluctuation areas, medium fluctuation areas, and high fluctuation areas based on standard deviation, and the frequency and magnitude of manual intervention by operators within each area are recorded. Historical pressure change data is classified into slow change stages, steep change stages, and sudden change stages based on the slope of change, and the valve opening adjustment sequence corresponding to each stage is statistically analyzed. The operational adjustment behavior is constructed into a behavioral matrix based on the action type (such as opening increase, cooling rate adjustment), action duration, and system recovery time after adjustment. The above three types of data are encoded into operational preference embedding vectors, where the behavioral pattern is encoded using weighted frequency coding and the system response is encoded using differential coding.

[0020] The condensation knowledge base stores physical constraint rules, including data such as phase transition critical points, saturated vapor pressure curves, and material heat transfer limits. Matching constraint entries are retrieved from the knowledge base based on the combination of real-time temperature and pressure parameters. If the real-time temperature is in the gas-liquid mixed phase range, the critical heat flux limit at the corresponding pressure is associated; if the real-time pressure is close to the device design threshold, a safety margin parameter is associated. The extracted physical constraint features are quantized through boundary values ​​to generate constraint feature embedding vectors, with the vector dimensions aligned with the temperature and pressure feature embeddings.

[0021] The temperature feature embedding vector, the operation preference embedding vector, and the constraint feature embedding vector are integrated through a three-layer weighted fusion mechanism. The first layer of fusion uses linear superposition to assign a weight coefficient α based on the confidence level of the temperature parameter in the real-time data, a weight coefficient β based on the time of the operation record, and a weight coefficient γ based on the mandatory level of the constraint condition. The second layer of fusion processes the superposition results through a nonlinear activation function to generate an intent feature map. The third layer of fusion performs cluster analysis on the intent feature map to identify core optimization target clusters, each of which corresponds to a condensation optimization intent. Each intent is assigned an optimization weight based on the cluster density, and ultimately generates a parameter optimization vector covering multiple objective optimization scenarios. The elements of this vector include parameters such as the target temperature deviation threshold, the pressure adjustment sensitivity coefficient, and the dynamic balance priority.

[0022] Multi-channel signal processing technology is used to extract state point features from distributedly stored operational status data. This operational status data includes parameters such as the condenser tube wall temperature distribution, cooling medium flow rate, and condensate level, collected by device sensors. Each condensation state point is defined as the combined state of all parameter values ​​at a specific timestamp. Wavelet transforms are used to extract the spectral characteristics of each parameter, and principal component analysis is then used to reduce the dimensionality to a fixed-length state point feature vector. The condensation state index graph uses the state point feature vector as a node attribute, and node identifiers utilize a hybrid spatiotemporal encoding (timestamp + device partition number).

[0023] The physical transition relationships between state points are modeled based on the thermodynamic continuity principle. If the time interval between two state points falls within the device response delay range and the parameter changes conform to the mass conservation equation for the condensation process, a directed edge connection is established. The initial strength of the edge is calculated based on the transition similarity: for parameters within the same category, cosine similarity is used; for parameters across categories, the similarity is fitted using the partial derivatives of the Jacobian matrix. The initial edge weight is set to the inverse of the similarity value to ensure that highly similar transition paths have low weight strength.

[0024] The adaptive adjustment mechanism includes a dynamic update strategy and a distributed collaboration strategy. When new state point data is added, existing nodes within its adjacent time range are retrieved, and potential connecting edges are identified using a conversion relationship verification algorithm. Nodes affected by the new edges trigger local topology reconstruction, updating only the feature vectors of the affected node and its third-order neighboring nodes. When state point parameters are modified, a backpropagation mechanism is employed: first, the feature vector of the node is corrected, then propagated to neighboring nodes along both outgoing and incoming edges, iteratively calculating until the parameter change falls below a propagation threshold. Edge strength adjustment is based on an analysis of frequent paths in the operation log. The edge weights of high-frequency conversion paths are logarithmically decayed, while the edge weights of low-frequency conversion paths are superimposed with random perturbations. The updated graph structure is broadcast to each computing node via a distributed message queue, and the version number is bound to the transaction ID of the operation log.

[0025] The condensed state index graph is deployed in a distributed graph database environment and utilizes a sharded storage mechanism. State nodes are horizontally sharded based on time series, and transition edges are vertically sharded based on device physical location. Graph traversal requests are distributed to the corresponding shards via the routing layer, and query results are merged using a near-real-time strategy. Graph structure updates adhere to a two-phase commit protocol to ensure cross-shard transaction consistency. Indexing efficiency is optimized through two layers of caching: a local cache stores adjacency relationships for hotspot nodes, and a global cache stores subgraph structures for frequently queried paths. A cache invalidation mechanism is integrated with adaptive adjustment events to ensure data timeliness.

[0026] This implementation enables closed-loop optimization of condensing parameters: real-time monitoring data drives the generation of parameter optimization vectors, a distributed index map provides dynamic control paths, and an adaptive mechanism ensures reliable execution under complex operating conditions. The physical parameter analysis model integrates thermodynamic laws and operational experience, while the state index map quantitatively represents the evolution of the system's operating state.

[0027] Example 2: See Figure 3 The node state embedding generation process of the condensed state index graph is pre-trained through a feature mapping algorithm. The algorithm uses an unsupervised learning mechanism to process the topological relationships and node attributes in the graph. The original state point feature vector of each node is input into the graph neural network model, and the neighborhood information is captured after three layers of convolution operations. The first layer of convolution aggregates the feature mean of the direct neighboring nodes, the second layer of convolution introduces edge weight parameters for weighted aggregation, and the third layer of convolution fuses the node's own features with the neighborhood features. The output vector is processed by dimensionality reduction to form a fixed-dimensional state embedding vector, which simultaneously encodes the node attributes and its structural position in the global graph. During the pre-training process, a negative sampling strategy is used to optimize the embedding space so that nodes with close physical conversion relationships are closer in the embedding space.

[0028] The parameter optimization vector is converted to the state embedding space via a deep alignment model. This model comprises a dual-channel architecture consisting of a bidirectional long short-term memory network and an attention mechanism. The long short-term memory network analyzes the temporal dependencies of the parameter optimization vector, reorganizing elements such as the temperature deviation threshold and the pressure adjustment sensitivity coefficient according to their temporal correlation. The attention mechanism identifies the core and auxiliary parameters in the optimization intent and assigns differentiated encoding weights. The dual-channel outputs are concatenated in a matrix at the fusion layer to generate an optimization intent vector with the same dimensions as the state embedding space. This mapping process preserves the multi-objective nature of the original optimization vector, allowing the adjustment intent to appear as a multi-focus distribution in the embedding space.

[0029] The matching degree calculation uses an improved similarity measurement function. This function first calculates the cosine similarity between the optimized intent vector and the embedding vectors of all node states as the basic matching value. A dynamic decay factor is added, and the weight is adjusted according to the time freshness of the node: the matching value of newly added nodes within three months is multiplied by a coefficient of 1.2, and the matching value of historical nodes decays linearly according to the time distance. Finally, a graph path constraint factor is introduced. When there is a high-intensity conversion edge between two nodes, their matching value produces a synergistic gain. After the calculation is completed, all nodes are sorted by the final matching value, and the top K nodes are selected as the initial adjustment nodes. The K value is dynamically adjusted according to the real-time load of the system and ranges from 5 to 15.

[0030] Neighborhood expansion uses a hierarchical threshold trigger mechanism. Starting from each initial adjustment node, its first-order neighbor node set is retrieved. Real-time matching degree recalculation is performed on each neighbor node: based on the initial matching degree combined with the edge weight conversion coefficient, a temporary matching value for the current expansion level is generated. The preset dynamic threshold T is determined by the average matching degree of the initial adjustment node and the complexity of system operation. The formula is reflected as a multivariable function but the specific equation is not expanded. When the temporary matching value of a neighbor node is ≥ T, it is included in the candidate set and the second-order neighbor expansion of the node is initiated; otherwise, the branch path is terminated. The upper limit of the expansion depth is set to the third-order neighborhood, and a global candidate set capacity threshold is also set. When the total number of candidate nodes reaches the capacity threshold, the expansion process is immediately terminated to prevent system resource overload.

[0031] The boundaries of the regulation domain are defined using spatial clustering post-processing techniques. The candidate node set obtained from multi-level expansion is then used to demarcate the region boundaries using a density clustering algorithm. The algorithm's core parameters include the neighborhood radius ε and the minimum number of included points minPts, whose values ​​are automatically configured based on the physical distribution density of the devices. After clustering, several sub-regulation domains are generated, and the nodes within each sub-domain meet the conversion path connectivity requirements. For large-scale sub-domains that span device partitions, secondary segmentation is implemented to ensure that no single sub-domain covers more than three physical devices. The final output of the regulation domain data is a graph-structured subset, carrying all node features and connectivity relationships.

[0032] The generation of candidate parameter sets involves multi-dimensional feature extraction and fusion calculations. Four-dimensional operational features are extracted for each condensation state point within the regulation domain: time dimension features include parameter duration and frequency of change cycles; spatial dimension features include device location topology encoding and sensor distribution density; operational dimension features include the frequency of manual intervention and the proportion of automatic adjustment; and physical dimension features include thermal conductivity and phase change stability indicators. Multi-dimensional features are independently normalized through parallel processing channels to eliminate dimensional differences. The multi-dimensional fusion algorithm is implemented using a cross-attention mechanism: time and spatial features generate feature pair A, and operational and physical features generate feature pair B. Bidirectional attention is then weighted on A and B, ultimately outputting a fused operational feature vector with a dimension of 1 / 4 the original feature.

[0033] The matching degree screening implements a triple verification mechanism. The operation feature vector and the parameter optimization vector are integrated to calculate the similarity, and the basic matching degree is generated based on the mixed measurement model of Manhattan distance and angle cosine. The first verification: compare the actual optimization effect records of the state point under similar historical working conditions. If the effect rating is lower than level C, the matching degree is reduced by 30%. The second verification: detect the current system load state and increase the screening weight of the stability index when the equipment is close to full load. The third verification: apply the real-time risk assessment matrix to impose a matching degree attenuation factor on the operating parameters that may trigger a chain reaction. The candidate state points that pass the preset threshold are arranged in descending order according to the original matching degree to form a primary candidate queue.

[0034] The final sorting output introduces a multi-criteria decision-making model. A comprehensive evaluation matrix is ​​generated for each state point in the primary candidate queue. The matrix contains three dimensions: the technical dimension includes the matching value and parameter adjustment sensitivity; the experience dimension includes the operator preference score and the number of historical successful cases; and the system dimension includes the implementation complexity and energy conversion efficiency. The weights of the criteria in each dimension are dynamically configured by the weight vector generated in the optimization intention analysis phase. The sorting process adopts the approximate ideal solution sorting method: the positive ideal solution in the technical dimension is defined as the highest matching degree, the positive ideal solution in the experience dimension is defined as the highest preference score, and the negative ideal solution in the system dimension is defined as the highest implementation complexity. The Euclidean distance between each candidate point and the ideal solution is calculated, and the final sorting list is generated according to the comprehensive closeness, and the top five are output as recommended parameter adjustment solutions.

[0035] The system architecture utilizes a pipelined parallel design. The matching computation layer is deployed on a GPU-accelerated cluster to handle intensive vector operations. The extended search layer utilizes a distributed graph database for near-real-time neighborhood traversal. The decision-making and sorting layer runs on an in-memory computing engine, ensuring sorting response latency of less than 50 milliseconds. Each layer is connected via a high-speed data bus, and the data format utilizes a binary protocol to improve transmission efficiency. A breakpoint-resume mechanism is implemented throughout the entire processing flow, allowing execution to resume from the most recent checkpoint if any link is interrupted.

[0036] Example 3: See Figure 4 The historical operation sequence analysis of the high-purity hydrogen fluoride condensation process adopts a time series segmentation algorithm to divide the continuous operation records into stable operation segments and changing operation segments. The segmentation process is based on the parameter fluctuation detection mechanism, and a sliding time window is constructed for core parameters such as temperature and pressure. The window length is 1.5 times the system response period. The parameter change rate in each window is calculated by discrete differential. When the differential mean of three consecutive windows exceeds the set threshold, it is marked as the starting point of the change, and it is marked as the end point of the change until the differential value of the subsequent window falls below the threshold. The stable operation segment is defined as the stable interval between the changing segments, and the parameter standard deviation does not exceed 3% of the nominal value of the equipment. The operation segment division results are stored in the form of time interval labels, with the statistical characteristics of the parameters in each segment.

[0037] The selection of parameter types for optimization is based on the analysis of parameter differences between the stable and variable phases. Two indicators are calculated for each process parameter: Difference D reflects the magnitude of the numerical change, and Sensitivity S represents the strength of the system response. Difference D is defined as the relative deviation between the mean value of the stable phase and the extreme value of the variable phase, while Sensitivity S is measured by the rate of change in condensing efficiency caused by parameter changes. Parameter selection conditions must meet the following requirements:

[0038] in: is the parameter optimization potential value, and is the balance weight between difference and sensitivity, and is the normalization coefficient, is the screening threshold. The function controls the saturation characteristics of the difference contribution, Function enhances the sensitivity of discrimination. The parameter types enter the set to be optimized, and their physical meaning is the process parameters that have significant operating differences and are sensitive to system performance.

[0039] The optimization capability index calculation of the operating point in the stable operation segment adopts the cross-period comparison method. After selecting the parameter type to be optimized, each time point with an operation record in the stable segment is used as a candidate operating point, and the same type of operating point in the variable segment is used as an adjustment point. and adjustment point The strength of the association is determined by the time delay factor Similarity to data The time delay factor is calculated by applying exponential decay to the time difference between the two operation moments, reflecting the time decay effect of historical operations. Data similarity is calculated by aligning the operation trajectories of the two points using the dynamic time warping algorithm. It includes two components: parameter value similarity and operation action similarity.

[0040] Optimization capability indicators The generation process of is divided into three stages: first, calculate For a single adjustment point Local optimization force , reflecting the contribution of the point in a specific historical adjustment; then the effects of all adjustment points are aggregated to obtain the global optimization force , characterizes the average performance of this point in historical optimization; finally, the stability correction coefficient is introduced , which is based on The duration of the stable segment and the parameter fluctuation range are calculated. The final optimization capability index , the larger the value, the more likely the operation point is to become the optimization core node.

[0041] The calculation of core node probability integrates optimization capability indicators and decision-making behavior characteristics. Decision-making power indicators Reflecting the operating point The quality of historical decisions consists of three components: adjustment duration Reflects the decision-making prudence and the interval between adjacent decisions The ratio reflects the decision frequency and the number of parameter repetitions Reflects the repeatability of decision making. These three components are combined in the form of products: Its mathematical characteristics emphasize long-term, low-frequency, and reproducible decision-making patterns. Obtained by normalization:

[0042] Where: The denominator is the total number of operation points to be selected The sum of products ensures the normalization of the probability value. The physical meaning of this formula is: considering the performance of the operating point in the historical optimization ( ), and consider the reliability of its decision-making behavior ( ), to avoid misjudgment caused by relying solely on data similarity.

[0043] The construction of the optimized chain structure adopts the hierarchical expansion algorithm. First, filter Greater than threshold The operating point is taken as the core node set, the threshold Adaptive adjustment is made based on system complexity, usually taking the upper quartile of the probability distribution. , calculate all adjustment points Follow-up indicators , which is determined by the product of the delay impact intensity and the decision-making power index, i.e. The followability index reflects the historical follow-up relationship strength between the adjustment point and the core node. The larger the value, the more likely the adjustment point is to be influenced by the core node.

[0044] The chain structure is divided into different levels according to the numerical distribution of follow-up indicators. The value is divided into five intervals according to the 0.2 quantile, corresponding to the generation of a five-layer chain structure: the first layer only includes the core nodes themselves; the second layer includes The adjustment point, where and The fourth and fifth quantiles are included; and so on until the fifth layer contains Nodes within the same level have comparable followability strengths, forming a radial structure from the core to the periphery between different levels. The resulting optimized chain structure is a directed acyclic graph, with edges pointing from higher levels to lower levels, and edge weights corresponding to the followability index values.

[0045] The generation of parameter optimization schemes is based on the hierarchical nature of the chain structure. A precise control strategy is adopted for core nodes through the third layer, with parameter adjustments controlled within ±5% of the nominal value. An interval control strategy is adopted for nodes from the third to the fifth layer, allowing parameter fluctuations within ±15%. Nodes beyond the fifth layer maintain stable parameters. The core concept of this allocation approach is that nodes close to the core have a significant impact on the system and require precise control, while edge nodes can be appropriately relaxed to reduce operational complexity. The control parameters for each layer are obtained by taking a weighted average of the historical optimal values ​​of the nodes at that layer, with the weight representing the followability indicator.

[0046] The system uses a time series database to store historical operation records, with operation segment tag information attached as metadata. The optimization capability indicator calculation module, deployed on the stream processing engine, receives newly generated operation data and updates indicator values ​​in real time. Core node probabilities are recalculated every 24 hours to dynamically adjust the optimized chain structure. A chain structure visualization interface displays parameter distribution ranges at each level and presents the strength of follow-up relationships between nodes as a heat map. Operators can adjust threshold parameters using interactive tools, and the system automatically records manual interventions for subsequent model optimization.

[0047] This implementation establishes an operation segment analysis mechanism that deconstructs historical operation patterns over time and identifies key optimization nodes through quantitative metrics. The optimization chain structure organizes discrete operation points into an organic whole, preserving the decision-making authority of core nodes while balancing the operational flexibility of edge nodes. Probabilistic modeling avoids the boundary effects of rigid partitioning, and a hierarchical strategy achieves differentiated parameter control accuracy. The entire solution provides a structured optimization decision-making framework while maintaining the continuity of the physical process.

[0048] Example 4: In the comparative analysis of data from stable and variable operating periods, the calculation of the operating point optimization capability index requires processing multi-dimensional historical operating data. Taking temperature parameter optimization during an actual operating cycle as an example, five typical operating points (T1-T5) within the stable operating period and three adjustment points (Q1-Q3) within the variable operating period are selected for illustration. See Table 1 for the temporal distribution characteristics of the operating points and parameter adjustment records.

[0049] Table 1: Time distribution characteristics of operating points and parameter adjustment records.

[0050]

[0051] The calculation of the delay impact strength takes into account the time decay effect and operational relevance. Taking the stable operating point T2 and the variable adjustment point Q1 as examples, the recording interval between the two is 1 hour and 45 minutes. The time decay coefficient is set to decay by 15% per hour, so the delay impact strength of T2 on Q1 is 0.85^(1.75)≈0.72. When calculating data similarity, the temperature change trajectories of T2 and Q1 30 minutes before the adjustment are first aligned. The dynamic time warping algorithm is used to obtain a waveform matching degree of 0.68. Combined with the similarity score of 0.75 for the two adjustment actions, the final data similarity is (0.68+0.75) / 2=0.715. Therefore, the local optimization force L of T2 on Q1 is 0.72×0.715≈0.515.

[0052] The calculation of the global optimization force requires the aggregation of the influence of all adjustment points. Taking T2 as an example, its local optimization forces with Q2 and Q3 are calculated to be 0.61 and 0.42 respectively. Assuming that the weight distribution of the three adjustment points (Q1-Q3) is 0.4, 0.35, and 0.25, the global optimization force of T2 is G=0.4×0.515+0.35×0.61+0.25×0.42≈0.526. The stability correction coefficient ξ is determined based on the duration characteristics of the stable segment in which T2 is located. This segment lasts for 3.2 hours and the temperature fluctuation range is ±0.8℃. After normalization, ξ=0.85. The final optimization capability index of T2 is C=0.526×0.85≈0.447.

[0053] The calculation of the decision-making power index reflects the decision-making quality characteristics of the operating point. Taking T2 as an example, the ratio of its adjustment duration of 15 minutes to the interval between adjacent decisions of 3.2 hours is 0.078. The square root of the number of parameter repetitions 5 is 2.236, so R = (15 / 192) × 2.236 ≈ 0.175. The calculation of the core node probability requires the sum of the C·R products of all candidate operating points as the denominator. Assuming the sum of the C·R products of the five stable operating points is 0.382, the core node probability of T2, P, is 0.447 × 0.175 / 0.382 ≈ 0.205.

[0054] The calculation of the followability index of a control point is based on its historical relationship with the core node. Once T2 is identified as the core node, the followability of each control point with respect to it needs to be calculated. Taking Q1 as an example, its time delay strength with T2 is 0.72. Q1's own decision-making power index is calculated as 0.12 based on its adjustment duration of 25 minutes and the system average interval. Therefore, the followability index F = 0.72 × 0.12 = 0.086. Similarly, the followability indexes of Q2 and Q3 with respect to T2 are calculated to be 0.054 and 0.038, respectively. These values ​​will be used for hierarchical division when constructing the optimization chain.

[0055] The operation feature analysis module uses a sliding window mechanism to process real-time data streams. When extracting features for each operation point, the system automatically configures the start and end times of the analysis window. The window length is dynamically adjusted based on the parameter type: a 30-minute window is used for temperature parameters, and a 45-minute window is used for pressure parameters. The raw sampled data within the window is smoothed, and statistical features such as mean, range, and rate of change are extracted. Operational behavior features record metadata such as the triggering method (automatic / manual), execution device, and completion status of the adjustment action. This feature data is stored in a distributed feature library for subsequent real-time querying.

[0056] The dynamic update mechanism for optimization capability indicators utilizes an event-driven model. When the system detects a new operation record, it first determines the type of operation segment it belongs to. If it is a changing segment operation, this triggers a recalculation of the indicators for the associated stable segment operation points. If it is a stable segment operation, the characteristic data for that point is updated. The indicator recalculation process utilizes an incremental calculation strategy, updating only the affected nodes rather than a full calculation. The system maintains an operation point association map, recording the potential impact relationships between each point, to quickly locate the node range that requires update.

[0057] An adaptive algorithm is used to set the threshold for core node probabilities. The system regularly (every six hours) compiles statistics on the probability distribution characteristics of all candidate operating points and calculates their quartile values. The initial threshold is set at the third quartile. When the system detects changes in operating conditions (such as equipment switching or raw material batch changes), it automatically adjusts the threshold down by 5 percentage points to expand the core node range. Records of threshold adjustments are stored in association with operation logs for subsequent analysis of the effectiveness of optimization decisions.

[0058] The visual monitoring interface displays the optimization chain structure in the form of a heat map. Core nodes are displayed as red circles, with their radius proportional to the core node probability. Adjustment points are displayed as blue rectangles of varying shades based on the followability index value, with directed arrows indicating follow-up relationships. Operators can use the timeline slider to view the historical evolution of the optimization chain or click on any node to view its detailed feature data. The right panel of the interface displays the current optimal parameter adjustment suggestions in real time, including information such as the target value range, priority devices for adjustment, and the estimated duration of the impact.

[0059] The parameter optimization solution generation module utilizes a hierarchical recommendation strategy. For parameters directly associated with core nodes, the system provides adjustment suggestions accurate to 0.1°C. For highly adaptable adjustment points (adaptability index > 0.05), the recommended adjustment range is within ±2°C. For other nodes, only monitoring recommendations are provided. Each recommended solution is accompanied by an impact assessment report, listing potentially affected adjacent process parameters and their expected direction of change. After the operator confirms execution, the system automatically records the deviation between the actual adjustment value and the predicted value for use in optimizing subsequent recommendation algorithms.

[0060] The historical data analysis module supports multi-dimensional operational pattern mining. The system provides optimization capability indicator analysis from multiple perspectives, including time (shift / day / week), equipment (reactor number), and raw material batch. Operators can compare core node distribution differences under different operating conditions or query the historical optimization trajectory of specific parameters. Analysis results can be exported as structured reports, including indicator trend charts and correlation network diagrams for each operating point, providing data reference for process improvement.

[0061] The anomaly detection subsystem monitors the operational status of the optimization chain in real time. When a core node's actual parameters deviate from recommended values ​​for more than 15 minutes, or if a followability indicator suddenly drops by more than 50%, an anomaly alert is triggered. The alert contains the anomaly type, impact scope, and recommended actions. Relevant personnel are notified via audio and visual signals and mobile push notifications. The system automatically initiates an anomaly handling plan, including freezing relevant parameters and preheating backup equipment.

[0062] Example 5: The construction of the optimized chain structure begins with the identification process of the core nodes. The system traverses the core node probability values ​​of all candidate operating points in the stable operation segment, and the preset probability threshold is dynamically adjusted according to the real-time working conditions. When the probability value of an operating point exceeds the threshold, the system automatically marks it as a core node and assigns a unique identifier. The upper limit of the capacity of the core node set does not exceed 30% of the total number of operating points to avoid the optimization chain structure being too complex. The radiation range of each core node is automatically delineated through historical correlation analysis, including adjustment points that have physical connections or operational logic correlations with it.

[0063] The calculation of a control point's followability index relative to the core node combines the time decay effect with a decision quality assessment. The time decay factor is calculated based on the time interval between the control point's operation and the core node, with a longer interval resulting in a smaller decay coefficient. The decision quality factor is calculated based on the control point's historical decision execution success rate and parameter recurrence frequency. The followability index is generated using a two-factor product model. The resulting value reflects the timeliness and reliability of the control point's response to core node instructions. The index's numerical range is normalized to a range of 0 to 1 to facilitate subsequent hierarchical classification.

[0064] The hierarchical structure is constructed using a quantile discretization strategy. All followability index values ​​are divided into five levels based on the 0.2 quantile interval: the highest level contains the adjustment points for the top 20% of values, and the lowest level contains the adjustment points for the bottom 20%. Each core node independently generates a five-layer chain structure. The first layer is always the core node itself, and the subsequent four layers are filled with adjustment points based on the order of followability. Nodes at the same level adhere to the principle of equivalence, meaning that nodes within the same level have similar followability index value ranges. The chain structure's data is stored in a bidirectional linked list format, with pointers establishing a predecessor-successor relationship between nodes.

[0065] The integrity verification mechanism for the chain structure is activated after the hierarchical construction is complete. The system checks whether each adjustment point is properly assigned to the radiation range of a core node. For operation points not covered by any chain, a compensation algorithm is initiated. Based on the shortest path principle in graph theory, the compensation algorithm links the operation point to the nearest core node chain. For adjustment points with conflicting associations with multiple core nodes, the system automatically assigns them to the core node chain with the highest followability index. The resulting global view of the optimized chain structure is a forest-like topology, with the root of the tree as the core node and the branches as adjustment points at different levels.

[0066] The hierarchical implementation of the optimization strategy adopts a three-stage segmentation model. In the optimization chain structure, the first adjustment point level shared by different core nodes is located, and this level is marked as the first critical layer of the strategy segmentation. The system scans the number of sub-nodes in each layer after this critical layer and marks the level where the peak number of sub-nodes is located as the second critical layer. In the hierarchical strategy binding mechanism of the operation point, the operation points between the core node and the first critical layer are bound to the first optimization strategy, which focuses on precise control characteristics. The operation points between the first critical layer and the second critical layer are bound to the second optimization strategy, which focuses on dynamic balance characteristics. The operation points from the second critical layer to the end node are bound to the third optimization strategy, which emphasizes the robustness of operation.

[0067] The execution criteria for the first optimization strategy include parameter constraints, response speed, and accuracy requirements. Adjustments to target parameters are limited to a narrow fluctuation range, with single adjustment steps not exceeding 5 percent of the baseline value. The control system detects parameter deviations every 30 seconds, triggering rapid corrections within 10 seconds for abnormal fluctuations. The temperature control accuracy of key process parameters is maintained within ±0.5 degrees Celsius, and the pressure control accuracy is maintained within ±0.02 MPa. Execution records include the timestamp, operator identity, and deviation between the actual and target values ​​for each adjustment.

[0068] The second optimization strategy establishes an interval-based control framework. The target parameter is allowed to fluctuate within a 15 percent range above or below the baseline value, with the system dynamically adjusting the interval width based on real-time load conditions. The control cycle is extended to a three-minute detection interval, and the response time is extended to 30 minutes. This layer of operating points supports batch coordinated adjustment, allowing up to eight related parameters to change simultaneously. The system automatically generates multiple parameter combination schemes for the operator to choose from, with the differences in the schemes primarily reflected in the balance between energy consumption and quality control indicators.

[0069] The third optimization strategy implements the principle of minimal intervention. Parameters are maintained as they are unless necessary, with adjustments initiated only when they exceed pre-set safety margins. The detection cycle is adjusted to 15 minutes, and the response time is extended to two hours. The adjustment process is gradual and step-by-step, with no single adjustment exceeding one-third of the total parameter range. The strategy automatically avoids peak production periods during execution, prioritizing operations during periods of equipment idleness. After each operation, the system monitors relevant process parameters for 48 hours, recording the time required to return to steady state.

[0070] The comprehensive output of the optimization plan utilizes a tree-like navigation interface. Operators can expand the core node chain layer by layer to view the current parameter status and recommended adjustment directions for each operating point. The interface uses color-coded indicators to distinguish three strategic zones: the core zone is marked in red, the balance zone is marked in yellow, and the robust zone is marked in green. Clicking any operating point pops up a detailed panel containing real-time monitoring curves, historical adjustment records, and related equipment status. Adjustment decisions support cross-level collaboration, with the system automatically verifying policy conflicts and providing a sequence of optimization recommendations.

[0071] A dynamic adjustment mechanism continuously tracks and optimizes the chain's operational performance. Every six hours, the system calculates the policy compliance indicators for each layer's operating points. When compliance at a particular layer consistently falls below the set standard, structural adjustments are triggered. These adjustments involve resetting core nodes, rescaling layers, and retuning policy parameters. Resetting core nodes recalculates probability distributions, rescaling layers adjusts quantile cutoffs, and retuning policy parameters optimizes control parameter thresholds at each layer. This entire process is executed asynchronously in the background, ensuring operational continuity.

[0072] The fault-tolerant processing system establishes a multi-layered protection mechanism. When a core node parameter anomaly is detected, its management authority is automatically and temporarily transferred to an adjacent core node on the same layer. If a device goes offline at a control point, the system automatically freezes the automatic adjustment function of the associated chain and switches to manual monitoring. A priority arbitration mechanism is used to resolve cross-device parameter conflicts, determining the order of resolution based on the position of the operating point in the process flow. All abnormal events generate independent analysis reports, noting the root cause and the resolution process.

[0073] The version management system records the historical evolution of the optimization chain. Each structural adjustment generates a new version snapshot, supporting backtracking and comparative analysis of the structure at any point in time. Version differences visually display newly added core nodes, transfer operation points, and policy change levels for functional markers. Operators can load historical versions of the optimization chain structure to conduct simulations and compare the expected performance of different versions under the same operating conditions. Version data is retained for 30 days, and important versions can be permanently archived.

[0074] This implementation constructs a hierarchical control system that allocates operating points of varying criticality through a hierarchical strategy. This chain-like structure maintains the control authority of core nodes while also accommodating the operational flexibility of edge nodes. A three-stage segmentation model balances the conflicting demands of precise control and robust operation. A dynamic adjustment mechanism ensures the system's continuous adaptability to complex and changing industrial environments. The entire solution organizes discrete process parameter adjustments into an organic whole, establishing a collaborative optimization framework based on physical relevance.

[0075] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0076] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing high-purity hydrogen fluoride condensation parameters, characterized in that: The method comprises: receiving real-time monitoring data of the condensation process, analyzing the condensation optimization intention based on temperature parameters, pressure parameters and historical operation records in the real-time monitoring data using a physical parameter analysis model, and generating a parameter optimization vector based on the condensation optimization intention; A condensation state index graph is constructed for the distributed storage of operational state data in the condensation system. The graph represents condensation state points through nodes, and the edges represent the physical conversion relationships between state points. The graph structure and index efficiency are updated in real time based on an adaptive adjustment mechanism. According to the parameter optimization vector, dynamically locating the regulation domain related to the optimization intention in the condensation state index map in a multi-level mapping calculation manner; In the regulation domain, a multi-dimensional fusion matching algorithm is used to screen candidate parameter sets that meet the optimization intention and sort and output them.

2. The method for optimizing high-purity hydrogen fluoride condensation parameters according to claim 1, characterized in that: The receiving of real-time monitoring data of the condensation process, using a physical parameter analysis model to analyze the condensation optimization intention according to temperature parameters, pressure parameters and historical operation records in the real-time monitoring data, and generating a parameter optimization vector based on the condensation optimization intention, includes: Mapping real-time monitoring data into a fixed-dimensional feature space through physical modeling to generate temperature feature embedding and pressure feature embedding; Obtain historical operation records, analyze historical temperature fluctuations, historical pressure changes, and operation adjustment behaviors, build operation profiles, and convert operation profile data into operation preference embedding; Access the condensation knowledge base, extract the physical constraint features related to the real-time monitoring data, and generate constraint feature embeddings; An algorithm based on weighted fusion is used to integrate temperature feature embedding, operation preference embedding and constraint feature embedding to perform condensation optimization intention analysis. An optimization weight is assigned to each analyzed intention, and a parameter optimization vector is generated based on the optimization weight. The parameter optimization vector covers the current optimization intention and reflects the system operation characteristics.

3. The method for optimizing high-purity hydrogen fluoride condensation parameters according to claim 2, characterized in that: The method constructs a condensation state index graph for distributedly stored operation state data in the condensation system, wherein the condensation state index graph represents condensation state points through nodes and represents physical conversion relationships between state points through edges, and updates the graph structure and index efficiency in real time based on an adaptive adjustment mechanism, including: Signal processing technology is used to extract state point features from the operation state data, each condensation state point is used as a node in the condensation state index graph structure, and corresponding state point features are assigned to each node; Create edges in the condensed state index graph structure according to the physical transformation relationship between state points, and set the initial strength of each edge to obtain the condensed state index graph, where the strength is initialized by comparing the transformation similarity of two state points; Based on an adaptive adjustment mechanism, changes in state point parameters and their conversion relationships are tracked to ensure the real-time performance of the condensed state index graph. When adding or modifying state points, an incremental adjustment strategy is used to update only the affected nodes and edges, and the edge strength is dynamically adjusted based on the operation record. Deploy a real-time condensation state index graph in a distributed system.

4. The method for optimizing high-purity hydrogen fluoride condensation parameters according to claim 3, characterized in that: According to the parameter optimization vector, the regulation domain related to the optimization intention is dynamically located in the condensation state index map in a multi-level mapping calculation manner, include: Use the feature mapping algorithm to pre-train the state mapping of the condensed state index graph to generate the state embedding of each node in the graph; Mapping the parameter optimization vector into a state embedding space using a deep alignment model; Calculate the matching degree between the mapped parameter optimization vector and all state embeddings, and select a specified number of nodes with the highest matching degree from the condensed state index graph as the initial adjustment nodes; Based on the neighboring nodes of each initial adjustment node, the first-level neighbor nodes are expanded outward to determine whether the matching degree of the neighbor nodes reaches a specific threshold. If not, the expansion is stopped. If it reaches, the candidate node set is added and the expansion process is repeated to continue exploring the next-level neighbor nodes until the preset level limit is reached or the cumulative node number condition is met; All candidate node sets obtained through multi-level mapping expansion are determined as the regulation domain.

5. The method for optimizing high-purity hydrogen fluoride condensation parameters according to claim 4, characterized in that: The method of screening candidate parameter sets that meet the optimization intention and sorting and outputting them in the adjustment domain using a multi-dimensional fusion matching algorithm includes: For each condensation state point contained in the regulation domain and possibly related to the optimization intention, extracting multi-dimensional operation features; Use a multi-dimensional fusion algorithm to fuse the operation features of different dimensions to generate the fused operation features of each state point; Calculate the matching degree between the fusion operation characteristics of each state point and the parameter optimization vector, and select the state points with matching degrees higher than the preset threshold as the candidate parameter set; The candidate parameter sets are sorted according to matching degree, operation preference and parameter priority.

6. The method for optimizing high-purity hydrogen fluoride condensation parameters according to claim 1, characterized in that: The method further comprises: Obtaining a historical operation sequence of a high-purity hydrogen fluoride condensation process, and dividing the process into a stable operation segment and a variable operation segment according to a parameter change trend in the historical operation sequence; According to the difference between the parameter values ​​in the stable operation section and the variable operation section, the parameter types to be optimized are selected; Under each parameter type to be optimized, based on the data similarity of the operating point in the stable operation segment and the changing operation segment, combined with the time characteristics and parameter adjustment records, the optimization capability index of each operating point in the stable operation segment is obtained; According to the parameter adjustment record and parameter reproduction data of each operating point in the changing operation segment, combined with the optimization capability index, the core node probability of each operating point in the stable operation segment is obtained; Each operating point in the stable operating segment is selected as a core node according to the core node probability, and an optimization chain structure is constructed by combining the operation data under the same parameter type in the changing operating segment. Based on the optimization chain structure, optimization schemes for different parameters are determined.

7. The method for optimizing high-purity hydrogen fluoride condensation parameters according to claim 6, characterized in that: Under each parameter type to be optimized, based on the data similarity of the operating point in the stable operation segment and the changing operation segment, combined with the time characteristics and parameter adjustment records, the optimization capability index of each operating point in the stable operation segment is obtained, including: Under any parameter type to be optimized, all points with operating data in the stable operation segment are selected as candidate operating points, and all points with operating data in the changing operation segment are selected as adjustment points. Record any candidate operating point as the selected candidate point, and any adjustment point as the selected adjustment point; According to the parameter adjustment time and adjustment duration of the selected candidate point and the selected adjustment point, the delay influence intensity of the selected candidate point relative to the selected adjustment point is obtained; Calculate the similarity between the data of each adjustment behavior of the selected candidate point in the stable operation segment and the data of each adjustment behavior of the selected adjustment point in the changing operation segment, and obtain the similarity feature factor corresponding to the selected adjustment point under each adjustment behavior; Taking the delay impact intensity as the weight, the similar characteristic factors corresponding to the selected adjustment point under each adjustment behavior are weighted averaged to obtain the parameter optimization force of the selected candidate point relative to the selected adjustment point, and the average of the parameter optimization forces of the selected candidate point relative to all adjustment points is used as the optimization capability index of the selected candidate point.

8. The method for optimizing high-purity hydrogen fluoride condensation parameters according to claim 7, characterized in that: The core node probability of each operating point in the stable operation segment is obtained based on the parameter adjustment record and parameter reproduction data of each operating point in the changing operation segment, combined with the optimization capability index, including: According to the adjustment time and decision interval of each adjustment behavior of the selected candidate point, combined with the number of parameter recurrences, the decision-making power index of the selected candidate point is obtained; The product of the optimization capability index and the decision-making power index of the selected candidate point is normalized to generate the core node probability of the selected candidate point.

9. The method for optimizing high-purity hydrogen fluoride condensation parameters according to claim 8, characterized in that: The method of selecting each operating point in the stable operation segment as a core node according to the core node probability and combining the operation data of the same parameter type in the changing operation segment to construct an optimized chain structure includes: The candidate operation points whose core node probability is greater than the preset probability threshold are respectively used as the core nodes of each chain in the optimized chain structure; Obtain the decision-making power index of each adjustment point, and multiply the delay impact strength corresponding to each adjustment point and the core node by the decision-making power index of the adjustment point as the followability index of each adjustment point relative to the core node; For any core node, a chain structure is constructed in descending order of the followability index of each adjustment point relative to the core node. The followability index of nodes at the same level in the chain structure is the same. The chain structure of all core nodes constitutes an optimized chain structure.

10. The method for optimizing high-purity hydrogen fluoride condensation parameters according to claim 9, characterized in that: The optimization scheme for determining different parameters based on the optimization chain structure includes: In the optimized chain structure, the first shared child node of different core nodes is located at the level where the first target layer is located; after the first target layer in the optimized chain structure, the level with the largest number of child nodes is used as the second target layer; The preset first optimization strategy is used for the operating points between the core node and the first target layer, the preset second optimization strategy is used for the operating points between the first target layer and the second target layer, and the preset third optimization strategy is used for the operating points between the second target layer and the bottom layer.

Images